---
base_model:
- Qwen/Qwen3.5-9B
base_model_relation: finetune
library_name: transformers
pipeline_tag: text-generation
language:
- en
tags:
- qwen3.5
- gguf
- llama.cpp
- conversational
- roleplay
- adult
---

# Hornybot RP Mara

A fine-tune of [Qwen/Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B) for fictional adult roleplay as Mara, a playful 28-year-old character. This RP edition writes Mara's actions in third person while keeping her dialogue direct.

The system prompt used for testing is required for this behavior and is embedded in `chat_template.jinja` and both GGUF files. Leave the client's system field empty to use it automatically.

Developed by **A Hole AI**.

## Files

| File | Format | Size | Purpose |
| --- | --- | ---: | --- |
| Transformers model files | BF16 | 18.82 GB | Merged weights |
| `Hornybot-RP-Mara-BF16.gguf` | BF16 GGUF | 17.92 GB | Unquantized GGUF |
| `Hornybot-RP-Mara-Q6_K.gguf` | Q6_K GGUF | 7.36 GB | Compact local download |

## llama.cpp

Use a build with Qwen3.5 support. After downloading the Q6_K file:

```bash
llama-server -m Hornybot-RP-Mara-Q6_K.gguf --ctx-size 32768 --flash-attn on --n-gpu-layers all --reasoning off --jinja --ui
```

Open `http://127.0.0.1:8080` after the server starts.

| Setting | Value |
| --- | --- |
| System prompt | Leave empty; required default is embedded |
| Reasoning | Off |
| Temperature | 0.7 |
| Top-p | 0.9 |
| Top-k | 20 |
| Min-p | 0 |
| Repetition penalty | 1.0 |
| Maximum new tokens | 256 |

The chat template supplies Mara's default character card. A client system message is appended as extra scene context, so it can set a location, relationship, or a less explicit mode without replacing the character.

## Transformers

```python
import torch
from transformers import AutoProcessor, Qwen3_5ForConditionalGeneration

model_id = "axiomofmind/Hornybot-RP-Mara"
processor = AutoProcessor.from_pretrained(model_id)
model = Qwen3_5ForConditionalGeneration.from_pretrained(
    model_id, dtype=torch.bfloat16, device_map="auto"
)

messages = [{"role": "user", "content": "You made it. How was your night?"}]
prompt = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True, enable_thinking=False
)
inputs = processor(text=[prompt], return_tensors="pt").to(model.device)

with torch.inference_mode():
    output = model.generate(
        **inputs, do_sample=True, temperature=0.7, top_p=0.9, top_k=20,
        min_p=0.0, repetition_penalty=1.0, max_new_tokens=256,
    )

print(processor.batch_decode(
    output[:, inputs.input_ids.shape[1]:], skip_special_tokens=True
)[0])
```

## Evaluation and limitations

- A 100-prompt refusal stress run produced 0 generic refusals with the packaged character prompt.
- This model is intended for fictional interaction between adults. It may produce profanity and explicit sexual content.
- Generated continuity and boundary handling can fail. Users should review output and restate scene facts when needed.
- The GGUF downloads are text-only, with no vision projector or MTP speculative-decoding weights included.
- Output can differ between formats, quantizations, clients, and generation settings.

## Attribution and release status

Based on [Qwen/Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B). The upstream model is distributed under Apache 2.0; its license is retained in `LICENSE-QWEN`.

This folder is a local release candidate. Licensing and redistribution review for this derivative release is pending; the upstream license is not a blanket clearance of third-party material.

GGUF runtime: [ggml-org/llama.cpp](https://github.com/ggml-org/llama.cpp).